5030 IEEE TRANSACTIONS ON SMART GRID, VOL. 12, NO.

6, NOVEMBER 2021

Chance Constrained Scheduling and Pricing for

Multi-Service Battery Energy Storage
Weifeng Zhong , Kan Xie , Yi Liu , Shengli Xie , Fellow, IEEE, and Lihua Xie , Fellow, IEEE

Abstract—This paper studies optimal day-ahead scheduling N OMENCLATURE

of grid-connected batteries that simultaneously provide three
services: 1) load shifting, 2) real-time balancing, and 3) primary
Indices and Sets
frequency control (PFC). The uncertainties of load and frequency i, I Index and set of batteries, i ∈ I.
are incorporated in the cost-minimizing scheduling problem via k Index of frequency time slots.
chance constraints. The resulting chance-constrained problem n Index of discharging or charging, n ∈ {d, c}.
is then reformulated into a mixed-integer second-order cone
program (MISOCP) that can be solved by commercial solvers. t, T Index and set of time periods, t ∈ T.
However, it is computationally formidable to obtain the globally
optimal solution to the MISOCP due to the big problem size. To Parameters
obtain a suboptimal solution quickly, a heuristic based on penalty
alternating direction method (PADM) is developed to solve the
e Vector that includes all uncertainties in the system.
MISOCP. Fixing the integer solution returned by the heuristic, fk Frequency deviation in slot k (Hz).
we adopt the duality of the second-order cone program (SOCP) to k Length of a frequency time slot (s).
price the three services in the local market. Theoretical analysis t Length of a time period (h).
of the market equilibrium, individual rationality, and balanced p , s Safety parameter for battery power/capacity.
budget is given. Real-world data of load, frequency, and price in
the French grid is used in simulation. The results show that the
ηid , ηic Discharging/charging efficiency of battery i.
proposed heuristic is computationally efficient, and the pricing σD,t Standard deviation of eD,t (kWh).
results can guarantee a positive utility for each of the batteries, σW,t Standard deviation of eW,t (kWh).
incentivizing them to provide services. σn Standard deviation of en (h).
Index Terms—Chance constraints, battery energy storage, ξn Equivalent discharging/charging time of a battery that
multi-service battery, primary frequency control, pricing. provides PFC in a time period (h), where n ∈ {d, c}.
Dt Predicted energy consumption in time t (kWh).
eD,t Random variable denoting the uncertainty of energy
consumption in time t (kWh).
eW,t Random variable denoting the uncertainty of RES
Manuscript received April 16, 2021; revised July 7, 2021; accepted August energy generation in time t (kWh).
16, 2021. Date of publication August 31, 2021; date of current version
October 21, 2021. This work was supported in part by the National Natural En Predicted value (i.e., mean) of ξn (h).
Science Foundation of China under Grant 62003099, Grant 61973087, and en Random variable denoting the uncertainty of ξn (h).
Grant U1911401; in part by the National Key R&D Program of China under fdb Deadband for PFC (Hz).
Grant 2020YFB1807805 and Grant 2020YFB1807800; and in part by the
State Key Laboratory of Synthetical Automation for Process Industries under fmax Full activation frequency deviation for PFC (Hz).
Grant 2020-KF-21-02. Paper no. TSG-00592-2021. (Corresponding author: Pmax
i Maximum discharging/charging power of battery i
Yi Liu.) (kW).
Weifeng Zhong is with the School of Automation, Guangdong University
of Technology, Guangzhou 510006, China, and also with the Guangdong Qbt Price at which energy is bought from the grid in time
Key Laboratory of IoT Information Technology, Guangzhou 510006, China t ($/kWh).
(e-mail: wfzhongs@gdut.edu.cn). Qst Price at which energy is sold to the grid in time t
Kan Xie is with the School of Automation, Guangdong University of
Technology, Guangzhou 510006, China, and also with the Key Laboratory ($/kWh).
of Intelligent Detection and Internet of Things in Manufacturing, Ministry of Qm Price of committed power for PFC ($/kW).
Education, Guangzhou 510006, China (e-mail: kxie@gdut.edu.cn). Simax Maximum energy state of battery i (kWh).
Yi Liu is with the School of Automation, Guangdong University of
Technology, Guangzhou 510006, China, and also with the Guangdong– Simin Minimum energy state of battery i (kWh).
Hong Kong–Macao Joint Laboratory for Smart Discrete Manufacturing, Wt Predicted RES energy generation in time t (kWh).
Guangzhou 510006, China (e-mail: yi.liu@gdut.edu.cn).
Shengli Xie is with the School of Automation, Guangdong University
of Technology, Guangzhou 510006, China, and also with the 111 Variables
Center for Intelligent Batch Manufacturing Based on IoT Technology, ϑn,i (·) Discharging/charging cost function of battery i ($),
Guangzhou 510006, China (e-mail: shlxie@gdut.edu.cn).
Lihua Xie is with the School of Electrical and Electronic Engineering, where n ∈ {d, c}.
Nanyang Technological University, Singapore 639798 (e-mail: ani,t Participation factor of the discharging/charging
elhxie@ntu.edu.sg). power of battery i for balancing in time t, where
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TSG.2021.3109140. n ∈ {d, c}.
Digital Object Identifier 10.1109/TSG.2021.3109140 gbt Energy bought from the grid in time t (kWh).
1949-3053 
c 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
ZHONG et al.: CHANCE CONSTRAINED SCHEDULING AND PRICING FOR MULTI-SERVICE BATTERY ENERGY STORAGE
gst Energy sold to the grid in time t (kWh).
pni,t Discharging/charging power of battery i for load
shifting in time t (kW), where n ∈ {d, c}.
pni,t (e) Discharging/charging power of battery i for load
shifting and balancing in time t (kW), where n ∈
{d, c}.
n
ri,t Discharging/charging power of battery i for PFC in
time t (kW), where n ∈ {d, c}.
rk Actual power for PFC in slot k (kW).
rm Committed power for PFC (kW).
si,t (e) Energy stored in battery i in time t considering
uncertainty e (kWh).
n
xi,t Binary variable indicating the discharging/charging
state of battery i in time t, where n ∈ {d, c}.
I. I NTRODUCTION
ATTERY energy storage characterized by high ramp
B rates and fast response can provide power systems with
multiple services such as 1) load shifting [1]: move energy
consumption and generation across time periods to prevent
network overload, reduce generation cost, and realize energy
arbitrage; 2) real-time balancing [2], [3]: offset unforeseen
real-time variations of consumption and generation to achieve
demand-supply balance; 3) frequency control [4]: respond to
fast frequency deviations for maintaining power quality and
system stability. In deregulated electricity markets, battery
operators would optimize the use of charging and discharging
power in multiple services to maximize revenue [5]. Via appro-
priate market and pricing rules, battery operators could be
incentivized to increase investment in high-performance bat-
teries. This will greatly benefit power systems in transitioning
to massive use of distributed, intermittent renewable energy
sources (RES) and minimum use of controllable thermal
generation.
This paper studies optimal day-ahead scheduling and pric-
ing for grid-connected batteries that simultaneously provide
three services: 1) load shifting, 2) real-time balancing, and
3) primary frequency control (PFC). The comparison with the
related works is summarized in Table I. Load shifting is the
most common battery service based on the foreseeable demand
and RES supply. In contrast, balancing and frequency control
require that part of battery power is reserved for responding
to real-time uncertain variations. In [8]–[11], power reserve
decisions are made based on the upper and lower bounds of
uncertainty. Their reserve results are mostly conservative, lead-
ing to relatively high system costs. Another flexible way to
handle uncertainty is to employ distributionally robust chance
constraints [2], [12]–[14], which have been increasingly used
in recent related works. Chance constraints are characterized
by the statistical information of uncertainty. Adjusting the
safety parameters in chance constraints can strike a balance
between reliability and economy [15]. In addition, most related
works [2], [6]–[14] focus on the scheduling for multi-service
batteries from a system-wide view but lack discussion on
profit allocation among multiple batteries that jointly provide
services. In this paper, we propose a pricing method for a local
Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
5031
market, which can quantify each battery’s income from each
service and incentivize batteries to provide services.
The technical contributions of this paper are as follows.
• We propose a chance-constrained model of batteries
incorporating load shifting, balancing, and PFC. In par-
ticular, empirical information of means and variances
is derived from real-world frequency deviation data,
characterizing the uncertainty of energy content in PFC.
• We reformulate the chance-constrained multi-service
scheduling problem to a mixed-integer second-order cone
program (MISOCP) that can be solved by off-the-shelf
solvers. However, getting the globally optimal solution
is badly time-consuming due to the big problem size.
We thus develop a heuristic based on penalty alternat-
ing direction method (PADM) [16] to solve the MISOCP.
The results show that the proposed heuristic can obtain
high-quality feasible solutions quickly.
• By fixing the integer solution returned by the heuris-
tic, the MISOCP becomes a second-order cone program
(SOCP). We then employ the duality of the SOCP to
price the three services in the local market. We also pro-
vide theoretical analysis of market equilibrium, individual
rationality, and balanced budget. The pricing results
show that each battery has a positive utility, i.e., service
incomes are enough to cover battery costs.
The novelty of this paper is twofold. 1) We consider multi-
period scheduling for batteries to provide three services, which
leads to a large-size mixed-integer problem, and then we newly
employ PADM as a heuristic to efficiently solve the problem.
2) We use the duality-based pricing method to determine the
profit allocation among batteries that cooperatively provide
multiple services. These two points are the major differences
compared with the existing works on multi-service energy
storage, e.g., [2], [6]–[14].
II. S YSTEM M ODEL
Consider a local system (e.g., a microgrid) consisting of
consumers, RES, and batteries. Each battery is indexed by
i ∈ I := {1, . . . , I}. The scheduling horizon is the next oper-
ating day which is divided into multiple time periods, and
each period is indexed by t ∈ T := {1, . . . , T}. Let Wt be
the predicted total energy (kWh) produced by RES and Dt be
the predicted total energy consumption (kWh) in t. We call
(Dt − Wt ) the (net) load of the system in t. Suppose that the
load and batteries are connected to a bus without considering
line flow limits. The data interaction among different entities
can be implemented by advanced communications and com-
puting technologies [17], [18]. In what follows, we introduce
the three battery services and the chance-constrained battery
model.
A. Load Shifting
Let pdi,t and pci,t denote the discharging and charging power
(kW) of battery i in t, respectively. Let gbt and gst be the
energy (kWh) that the system buys from and sells to the grid,
respectively. The energy balance of the system for t ∈ T is
5032 IEEE TRANSACTIONS ON SMART GRID, VOL. 12, NO. 6, NOVEMBER 2021

TABLE I
R ELATED W ORKS ON M ULTI -S ERVICE E NERGY S TORAGE

described by C. Primary Frequency Control (PFC)

  The third battery service is to provide the grid with PFC.
pdi,t − pci,t + gbt − gst = Dt − Wt , (1) Let rm ≥ 0 be the power (kW) committed by the local system
i∈I
for PFC. Once rm is determined, the system should ensure that
pdi,t ≥ 0, pci,t ≥ 0, gbt ≥ 0, gst ≥ 0. (2) ±rm is always available for all periods t ∈ T on the operating
day. The committed power ±rm is jointly provided by multiple
In this paper, we set the length of a time period to t = 1 h, so d and r c denote the discharging and charging
batteries. Let ri,t i,t
t is omitted in the first term in (1). The first battery service power (kW) committed by battery i for PFC in t, respectively.
is to shift the system’s load among periods in order to reduce We have
the cost in trading energy with the grid. For this service, the  
decision variables are g := {gbt , gst }t∈T and p := {pi }i∈I where d
ri,t = rm , c
ri,t = rm , (8)
pi := {pdi,t , pci,t }t∈T . i∈I i∈I
rm ≥ 0, d
ri,t ≥ 0, c
ri,t ≥ 0. (9)
B. Real-Time Balancing
The decision variables for PFC are r := {rm , {ri }i∈I } where
Dt and Wt are predictive values that are rarely equal to their ri := {ri,t
d
, ri,t
c
}t∈T . Note that r denotes the committed (maxi-
actual measurements. Define eD,t as a random variable repre- mum) power. On the operating day, the actual power for PFC
senting the consumption uncertainty (kWh) with a mean of will vary according to real-time frequency deviation.
E[eD,t ] = 0 and a variance of Var[eD,t ] = σD,t 2 . Similarly,
Next, we model the uncertainty of PFC in the time domain.
let eW,t stand for the RES uncertainty with E[eW,t ] = 0 A frequency time slot is indexed by k. The length of a
and Var[eW,t ] = σW,t 2 . Let e be a vector representing all the
frequency time slot is denoted by k (e.g., 10 s) that is
uncertainties in the system. Let pdi,t (e) and pci,t (e) denote the much shorter than the length of a time period t = 1 h.
discharging and charging power under uncertainty e, respec- The frequency deviation (Hz) is given by fk = fN − fk where
tively. The energy balance under uncertainty for t ∈ T is fN is the nominal frequency (e.g., 50 Hz) and fk is the local
described by frequency measured at time slot k. Put aside the batteries
    indexed by i ∈ I temporarily and consider a conceptual bat-
pdi,t (e) − pci,t (e) + gbt − gst = Dt + eD,t − Wt + eW,t , tery that fully provides the committed power ±rm . Define rk
i∈I as the actual power of the conceptual battery responding to
(3) frequency deviation fk with −rm ≤ rk ≤ rm . Following the
 
pdi,t (e) = pdi,t + adi,t eD,t − eW,t , (4) proportional rule of PFC [4], we have
  ⎧
pci,t (e) = pci,t − aci,t eD,t − eW,t . (5) ⎨ 0, if |fk | ≤ fdb ;
rk = rm (fk /fmax ), if fdb < |fk | ≤ fmax ; (10)
In (4) and (5), {pdi,t , pci,t } are scheduled values, and {adi,t , aci,t } ⎩
rm (fk /|fk |), if |fk | > fmax .
are the participation factors of battery i in t. A higher adi,t
(or aci,t ) means that battery i reserves more discharging (or In the first case in (10), fdb ≥ 0 denotes the deadband.
charging) power for load uncertainty (eD,t − eW,t ). We define Any frequency deviation within the deadband will not be
  responded. In the second case, fmax > 0 denotes the full acti-
adi,t + aci,t = 1, (6) vation frequency deviation, and rk is proportional to fk . In
i∈I the third case, rk is at its maximum or minimum, ±rm . Here,
adi,t ≥ 0, aci,t ≥ 0. (7) {fdb , fmax } are fixed parameters that are generally decided
by the grid. The grid could also provide several choices of
The second battery service is to reserve batteries’ power to {fdb , fmax }, and the local system chooses one of them.
offset the realization of load uncertainty. The corresponding By the second and third cases in (10), it can be seen that
decision variables are a := {ai }i∈I where ai := {adi,t , aci,t }t∈T . the conceptual battery is discharged (rk > 0) if fk > 0, and
Once {p, a} are determined and the realization of (eD,t − eW,t ) it is charged (rk < 0) if fk < 0. In a period of t , we
is known, the actual battery operation on the operating day define two sets of time slots: K d1 = {k | fdb < fk ≤ fmax }
can be directly provided by (4) and (5). and K d2 = {k | fk > fmax }. These two sets are composed of

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
ZHONG et al.: CHANCE CONSTRAINED SCHEDULING AND PRICING FOR MULTI-SERVICE BATTERY ENERGY STORAGE 5033

 

t
1 d
+ ηic ri,τ
c
(Ec + ec ) − d ri,τ (Ed + ed ) , (12)
τ =1
ηi
where the first square bracket denotes the energy change
caused by load shifting and balancing, and the second square
bracket is the energy change caused by PFC.
Next, we define the chance constraints for pdi,t (e), pci,t (e),
and si,t+1 (e). Without assuming that e follows a certain distri-
Fig. 1. Histograms of the equivalent discharging time ξd . fk is given by bution, we consider that the underlying distribution of e comes
the 10-second frequency deviation data (k = 10 s) measured during Dec. from a family of distributions (i.e., an ambiguity set) that share
2019 in the French grid [19]. the same mean and covariance matrix [15]. According to (4)
and (5), we define a random vector ep,t := {eD,t , eW,t } rep-
resenting the uncertainty relating to {pdi,t (e), pci,t (e)}. Let p,t
the time slots at which the battery is discharged. According be the covariance matrix of ep,t . The ambiguity set for ep,t is
to (10), the discharging energy (kWh) in t is denoted by defined as
  fk      

rk k = rm k + K d2 rm k p,t := P | EP ep,t = 0, EP ep,t ep,t = p,t , (13)
f max
k∈K d1 ∪K d2 k∈K d1
⎛ ⎞ where P stands for a probability distribution/function. The
chance constraints for discharging and charging power of
⎜  fk ⎟ battery i ∈ I in t ∈ T are written as
= rm k⎝ + K d2 ⎠
d
f max  
k∈K 1 inf P 0 ≤ pdi,t (e) + ri,t
d
≤ xi,t
d max
Pi ≥ 1 − p , (14)
P∈ p,t
= rm ξd , (11)  
inf P 0 ≤ pci,t (e) + ri,t
c
≤ xi,t
c max
Pi ≥ 1 − p , (15)
where ξd is called the equivalent discharging time (h) of a P∈ p,t

battery that provides PFC in t. By (11), the uncertainty in where Pmax is the maximum discharging/charging power, and
i
PFC is transferred from rk to ξd that is in the time domain. {xi,t
d
, xi,t
c
} are binary variables satisfying
Similarly, if battery i has a committed discharging power of
d in time period t, its discharging energy caused by PFC is
ri,t
d
xi,t , xi,t
c
∈ {0, 1}, d
xi,t + xi,t
c
= 1, (16)
d ξ . Following the same way, we can define the equivalent
ri,t d which means that battery i is either discharged or charged
charging time ξc . in t. Define x := {xi }i∈I where xi := {xi,t d
, xi,t
c
}t∈T . In (14)
According to the definition in (11), ξd is determined by and (15), p > 0 is the safety parameter, a small constant.
predefined parameters {k, fdb , fmax } and uncertain devia- A smaller p implies a smaller chance to violate the power
tions fk . Here, we consider ξd as a random variable with bounds [0, Pmax ].
i
E[ξd ] = Ed and Var[ξd ] = σd2 . Fig. 1 shows the histograms of Similarly, from (12), we define a vector es,t := {eD,1 , . . . ,
ξd [19]. Let ξd = Ed + ed where ed represents the uncertainty eD,t , eW,1 , . . . , eW,t , ed , ec } representing the uncertainty relating
of ξd with E[ed ] = 0 and Var[ed ] = σd2 . Hereafter, we use to si,t+1 (e), with a covariance matrix denoted as s,t . Similar
{Ed , ed } instead of ξd . Similarly, we define Ec as the mean of to (13), define an ambiguity set s,t with respect to es,t . The
ξc and ec as the corresponding uncertainty with E[ec ] = 0 and chance constraint for si,t+1 (e) for i ∈ I, t ∈ T is given by
Var[ec ] = σc2 . In simulation, {Ed , Ec , σd2 , σc2 } are sampling  
values derived from the French grid data [19]. Now, we have inf P Simin ≤ si,t+1 (e) ≤ Simax ≥ 1 − s , (17)
P∈
d
ri,t (Ed + ed ) and ri,t
c
(Ec + ec ), which denote the ith battery’s s,t

discharging and charging energy (kWh) caused by PFC in time
period t, respectively. These two terms relate to battery energy
state evolution and operating cost, which will be presented in
the following sections. In this paper, the uncertainty of PFC
is described by {ed , ec }.
D. Chance-Constrained Battery Model
The system uncertainty vector can be formally defined
as e := {{eD,t , eW,t }t∈T , ed , ec } including the uncertainties of
demand, RES, and PFC. Let ηid and ηic be the discharging and
charging efficiencies of battery i, respectively. Define si,t (e) as
the energy (kWh) stored in battery i in t under uncertainty e.
The energy state evolution of battery i is given by
 
 t
1 d
si,t+1 (e) = si,1 + ηi pi,τ (e) − d pi,τ (e)
c c
τ =1
ηi
where and Simax are the minimum and maximum energy
Simin
states of battery i, respectively, and s > 0 indicates the
chance to violate the energy bounds [Simin , Simax ]. Recall that
the uncertainty of PFC is modeled in the time domain in the
previous subsection. Thus, the power chance constraints (14)
and (15) only include the uncertainty of balancing, while the
energy chance constraint (17) includes the uncertainties of
both balancing and PFC.
III. P ROBLEM F ORMULATION
A. System Cost Minimization
Let Qbt and Qst be the prices ($/kWh) at which the system
buys energy from and sells energy to the grid, respectively.
The cost of trading energy with the grid in t is given by
Qbt gbt − Qst gst . Let Qm be the price ($/kW) of providing PFC
to the grid for one day. The corresponding income is Qm rm .

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
5034 IEEE TRANSACTIONS ON SMART GRID, VOL. 12, NO. 6, NOVEMBER 2021

Quadratic functions [15], [20] are used to quantify the bat- 3) Chance Constraints: As per [15], a chance constraint in
tery discharging and charging costs: ϑn,i (χ ) = Cn2,i χ 2 + the form of
Cn1,i χ , ∀n ∈ {d, c}, where {Cn2,i , Cn1,i } are constants. The  
inf P v e + w ≤ U ≥ 1 −  (21)
system cost minimization problem is given by P∈
  can be exactly reformulated as a set of deterministic
min Qbt gbt − Qst gst − Qm rm inequalities:
g,p,a,
r,x t∈T
    y2 + v v ≤ (U − z)2 , |w| ≤ y + z,
+ E ϑn,i pni,t (e) + ri,t
n
(En + en ) , (18)
n∈{d,c}, 0 ≤ y, 0 ≤ z ≤ U, (22)
i∈I,t∈T
s.t. (1), (2), (4)–(9), (12), (14)–(17), where {y, z} are auxiliary variables; denotes the covariance
matrix of the random vector e. In (21), is the ambigu-
where a new index n ∈ {d, c} is used to represent discharg- ity set defined with respect to e, similar to (13). Here, we
ing or charging. It can be found that (3) is equivalent to (1) first write the chance constraints (14), (15), (17) in the form
if (4)–(6) hold, so (3) is omitted here. It is difficult to directly of (21), and then transform them into the form of (22).
solve problem (18) due to the random variables e and chance Define auxiliary variables {y, z} := {yi , zi }i∈I where {yi , zi } :=
constraints (14), (15), (17). In what follows, we reformulate {yni,t , zni,t }n∈{d,c,s},t∈T . We have
problem (18) to a deterministic form.  n 2  2  
2  n 2
yi,t + σD,t + σW,t 2
ani,t ≤ p Ui,t − zni,t , (23a)

B. Problem Reformulation pni,t − Ui,t

n
≤ yni,t + zni,t , (23b)
 n max 
1) Objective Function: To simplify the problem, we assume 0≤ yni,t , 0≤ zni,t ≤ n
Ui,t := xi,t Pi − ri,t /2, (23c)
n

that the entries in e are uncorrelated with each other. This where n ∈ {d, c} corresponding to (14) and (15), respectively.
assumption is acceptable if the uncertainty in the system is We have
mostly caused by discrete and changeable factors. For exam-  s 2   2
ple, demand uncertainties {eD,t }t∈T are generally uncorrelated yi,t + Mi,t
s
≤ s Simax − Simin /2 − zsi,t , (24a)
if they are caused by the random use of, say, elevators, Λsi,t ≤ ysi,t + zsi,t , (24b)
motion sensor lights, and automatic doors. The uncorrelation  max 
0≤ ysi,t ,0≤ ≤ Si − Si
zsi,t min
/2, (24c)
assumption is often used in related works [2], [13], [14] for  2
model/problem simplification. By this assumption, we have t  1
s
Mi,t := σD,τ
2
+ σW,τ
2
ηic aci,τ + d adi,τ
Var[eD,t − eW,t ] = σD,t2 + σ 2 and E[e e ] = E[e e ] = 0
W,t D,t n W,t n η
τ =1 i
where n ∈ {d, c}. Then, the expectation in (18) can be  2  2
expressed in a deterministic form:  t
1  t
+ σc2 ηic c
ri,τ + σd2 d d
ri,τ , (24d)
   τ =1
η i τ =1
E ϑn,i pni,t (e) + ri,t
n
(En + en )  
 2     t
  1 d 
2
= Cn2,i pni,t + Cn2,i σD,t 2
+ σW,t2
ani,t Λi,t :=
s
ηi pi,τ + Ec ri,τ − d pi,τ + Ed ri,τ
c c c d
   τ =1
ηi
+ Cn2,i En2 + σn2 ri,t n 2
+ 2Cn2,i En pni,t ri,t
n  max 
+ si,1 − Si + Simin /2, (24e)
+ Cn1,i pni,t + Cn1,i En ri,t
n
, (19)
corresponding to (17). Note that (23a) and (24a) are second-
which is a nonconvex function due to the bilinear term pni,t ri,t
n . order cone (SOC) constraints because of the uncorrelation
2) Convex Envelopes: We eliminate the bilinear terms by assumption.
relaxing them to their convex envelopes, similar to [21]. A
bilinear term pr with Pl ≤ p ≤ Pu and Rl ≤ r ≤ Ru can be C. Mixed-Integer Second-Order Cone Program (MISOCP)
relaxed to a convex envelope h defined as By taking pni,t ri,t
n as an individual variable, we write

the battery cost function (19) as ϕi,t n (pn , an , r n , pn r n ).

i,t i,t i,t i,t i,t
h ≥ Pl r + pRl − Pl Rl , h ≥ Pu r + pRu − Pu Ru , Problem (18) can be transformed to the following MISOCP:
h ≤ Pl r + pRu − Pl Ru , h ≤ Pu r + pRl − Pu Rl ,  
min Qbt gbt − Qst gst − Qm rm
g,p,a,r,x,
y,z,pr t∈T
which is known as McCormick envelopes [22]. In our model, 
we have 0 ≤ pni,t , ri,t
n ≤ Pmax . Thus, the bilinear term pn r n + ϕi,t
n
(pni,t , ani,t , ri,t
n
, pni,t ri,t
n
), (25)
i i,t i,t
can be relaxed to n∈{d,c},
i∈I,t∈T
 n   2 s.t. (1), (2), (6)–(9), (16), (20), (23), (24).
pni,t ri,t
n
≥ 0, pni,t ri,t
n
≥ Pmax
i ri,t + pni,t − Pmax
i , (20a)
pi,t ri,t ≤ Pi pi,t , pi,t ri,t ≤ Pi ri,t ,
n n max n n n max n
(20b) This problem is a relaxation of problem (18) due to (20),
but the impact of the relaxation on the optimal objective
where we take pni,t ri,t
n as a new individual variable. Define pr :=
value would be small since En is generally small in prac-
{pri }i∈I where pri := {pni,t ri,t
n }
n∈{d,c},t∈T . tice (En ≈ 0.066 h in simulation). Note that the relaxation

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
ZHONG et al.: CHANCE CONSTRAINED SCHEDULING AND PRICING FOR MULTI-SERVICE BATTERY ENERGY STORAGE 5035

only applies to the objective function, so any feasible solution Algorithm 1: Heuristic Based on PADM
to problem (25) is still feasible to the original problem (18). 1 Input: m = 0, {Y 0 , Z0 }, ρ > 0, ρ > 0, α = 1, 0 < θ < 1.
Constraints gbt gst = 0 and adi,t aci,t = 0 are not included 2 Output: Integer solution Z∗
√
in problem (25) as they can be satisfied automatically (see 3 Normalization factor: W = TI ∇f (Y 0 ) −1 2 .
Appendix A). 4 repeat // Outer loop
MISOCP is mostly NP-hard and can be solved by exact 5 repeat // Inner loop
algorithms (e.g., branch and bound) via off-the-shelf solvers 6 Continuous solution: Y m+1
(e.g., CPLEX). See [23] for a review of the algorithms = arg min {φ(Y, Zm ; ρ, α) | Y ∈ Y}. (27)
and solvers for convex mixed-integer nonlinear programs Y
(MINLP). However, if the time spent by exact algorithms to 7 Integer solution: Zm+1
get the optimal solution is too long, we may resort to heuris-
tics. For instance, we set up problem (25) with 3 batteries and = arg min {φ(Y m+1 , Z; ρ, α) | Z ∈ Z}. (28)
Z
24 time periods, and CPLEX was unable to prove the opti-
8 Inner loop index: m ← m + 1.
mality of the solution in 10 hours. For practicality, a heuristic
9 until Zm = Zm−1 ;
is developed in the next section to get suboptimal solutions 10 Weight parameter update: α ← θα.
quickly. 11 Penalty parameter update:
d,m
0 ← ρ 0 + ρ,
ρi,t i,t if Zi,t = 0. (29)
IV. S OLUTION 1 ← ρ 1 + ρ,
ρi,t d,m
if Zi,t = 1. (30)
i,t
A. Penalty Alternating Direction Method (PADM) m m
12 until x = Z ;
Based on PADM [16], we develop a heuristic to obtain
a suboptimal solution to problem (25) quickly. First, we
transform problem (25) to make it suitable for the PADM
d,m
framework. Let the objective of problem (25) be written as Due to Zi,t ∈ {0, 1}, the objective in (27) can be written as
minY f (Y), and let Y ∈ Y denote the continuous relaxation 
n ∈ {0, 1} is relaxed to 0 ≤
of the constraints in (25), i.e., xi,t min αWf (Y) + (1 − α) ρi,t
0 d
xi,t
Y
xi,t ≤ 1. Define auxiliary variables Z := {Zi,t
n n }
n∈{d,c},i∈I,t∈T d,m
{i,t}:Zi,t =0
and an integer set Z := {Z | Zi,t ∈ {0, 1}, Zi,t
n d + Z c = 1}.   
i,t + (1 − α) ρi,t
1
1 − xi,t
d
. (31)
Problem (25) can be written in an equivalent form:
d,m
{i,t}:Zi,t =1
min f (Y), s.t. Y ∈ Y, Z ∈ Z, x = Z,
Y,Z This shows that problem (27) is a convex program. In
problem (28), f (Y m+1 ) is a constant, so the solution Zm+1
where x is part of Y. By the definitions of x and Z, we have can be directly given by
|xi,t
d − Z d | = |xc − Z c |. This implies that we can simplify the
i,t i,t i,t   
constraint x = Z to either xi,td = Z d or xc = Z c . Here we
i,t i,t i,t d,m+1
0 d,m+1
0, if ρi,t xi,t 1 1 − xd,m+1 ;
≤ ρi,t
choose the former, which is further decomposed into Zi,t = i,t (32)
1, otherwise.
d
xi,t ≥ Zi,t
d
, d
xi,t ≤ Zi,t
d
. In the outer loop, weight parameter α and penalty param-
eters ρ are updated. The purpose of updating α is to obtain
Then, we penalize them in the objective function. The resulting a feasible solution. α will gradually decrease to increasingly
penalty problem is given by emphasize the penalty terms so as to reach x = Z. The purpose
   of updating ρ is to obtain a good integer solution Z. The idea
d +
min αWf (Y) + (1 − α) ρi,t
0 d
xi,t − Zi,t is to use perturbation that allows the algorithm to try different
Y,Z
i∈I,t∈T choices of integers. Here, the perturbation is implemented by
 d d +
 
+ ρi,t
1
Zi,t − xi,t , (26) the penalty parameter update in (29) and (30). Fig. 2 shows the
d d
effect of (29) and (30) on integer solution Zi,t if xi,t is fixed. It
s.t. Y ∈ Y, Z ∈ Z,
can be seen that if xi,t is far from both 0 and 1 (xi,t = 0.5, 0.7),
d d
d
where [χ ]+ means max{0, χ }; ρ := {ρi,t 0 , ρ1 } then Zi,t is frequently shifted between 0 and 1. This can be
i,t i∈I,t∈T are
penalty parameters; 0 ≤ α ≤ 1 is a weight parameter for understood as strong perturbation that lets the algorithm keep
trying between Zi,t d = 0 and Z d = 1. Fig. 2 also shows that
balancing the system cost f and penalty; W is a normalization d
i,t
factor [16]. if xi,t is closer to 0 or 1, then less perturbation appears. This
can assist the algorithm in converging to xi,t d ≈ Z d . Fig. 2 is
Next, we employ PADM to solve problem (26), as shown i,t
in Algorithm 1. The objective function in (26) is written as only used for demonstrating the effect of the penalty parameter
d is given by the solution to problem (27), so
φ(Y, Z; ρ, α). The algorithm is composed of two loops. The update. In fact, xi,t
d
xi,t is generally varying (not fixed) in the algorithm. Therefore,
inner loop, indexed by m, is a standard ADM that can converge
to partial minima of problem (26) when {ρ, α} are given [16]. the perturbation frequency can adapt to xi,t d , helping the inner

The two problems (27) and (28) can be simplified as follows. loop reach xi,t ≈ Zi,t .
d d

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
5036 IEEE TRANSACTIONS ON SMART GRID, VOL. 12, NO. 6, NOVEMBER 2021

which the batteries and load trade energy in the system

at t. Define λ := {λt }t∈T .
2) For balancing, define μt as the multiplier of con-
straint (6). It is the payment ($) made by the load for
buying balancing reserve in t. Define μ := {μt }t∈T .
3) For PFC, define πtd and πtc as the multipliers of the
two constraints in (8), respectively. They are the prices
($/kW) of discharging and charging power reserved for
PFC in t. Define π := {πtd , πtc }t∈T .
With the prices  := {λ, μ, π }, we analyze the system from
a market perspective. Here, we consider a local market with
two types of agents: batteries and a system manager. For bat-
Fig. 2. The effect of the penalty parameter update on integer solution Zi,t d tery i ∈ I, its individual utility maximization problem is given
in the case that xi,t is fixed and ρ = 10. The x-axes denote the iteration of
d by
the outer loop in Algorithm 1.      
max λt pdi,t −pci,t + μt adi,t +aci,t + πtd ri,t
d
+ πtc ri,t
c
pi ,ai ,ri ,
yi ,zi ,pri t∈T

In summary, the inner loop of Algorithm 1 is responsible − ϕi,t
n
(pni,t , ani,t , ri,t
n
, pni,t ri,t
n
), (34)
for providing a partial minimum of penalty problem (26) with t∈T,
given ρ and α. Note that the partial minima satisfy Y ∈ Y and n∈{d,c}
Z ∈ Z. In the outer loop, the update of ρ helps the inner loop s.t. pni,t ≥ 0, ani,t
≥ 0, ri,tn
≥ 0, ∀n ∈ {d, c}, t ∈ T,
explore different integer solutions and converge to x ≈ Z. The Relaxation constraints: (20), ∀n ∈ {d, c}, t ∈ T,
update of α gradually emphasizes the penalty terms in (26)
Power constraints: (23), ∀n ∈ {d, c}, t ∈ T,
for reaching x = Z. Therefore, if Algorithm 1 terminates, it
can return a feasible solution to the original problem (25). Energy constraints: (24), ∀t ∈ T.
In this problem, x is fixed to Z∗ . As shown by the objec-
B. Implementation tive function (34), a battery’s utility is defined as the incomes
According to (8), if the system provides PFC (rm > 0), it minus the operating cost. The system manager is a non-profit
d > 0 and 
i∈I i,t > 0 for every
should ensure both i∈I ri,t r c agent, acting as an interface between the grid and batteries.
t ∈ T. This means that the system needs at least one battery The system manager’s problem is given by
d
with xi,t = 1 and at least one battery with xj,t c
= 1, j = i for   
max Qm rm − πtd ri,t
d
+ πtc ri,t
c
every
 t. But, the integer solution given by (32) often yields g,rm
Z d = 0 or  Z d = I for some t, which results in 
i∈I t∈T

i∈I i,t i∈I i,t
rm = 0, i.e., no PFC service. To guarantee the availability of + Qst gst − Qbt gbt
t∈T
PFC, we can add the following constraint to problem (28).  
   
1≤ d
Zi,t ≤ I − 1. (33) + λt pi,t − pi,t + Dt − Wt ,
c d
(35)
i∈I t∈T i∈I
s.t. rm ≥ 0, gbt ≥ 0, gst ≥ 0, ∀t ∈ T.
When Algorithm 1 starts, Y 0 can be given by the solu-
tion to {minY f (Y) | Y ∈ Y}, while Z0 can be randomly The objective (35) is defined as the PFC payment from the
generated (because of the initial value of α = 1). For imple- grid minus the PFC payments to batteries plus the energy pay-
mentation, we change the outer loop stopping condition to ments from the grid, batteries, and load. The market clearing
xm − Zm ∞ < 10−3 and xm − xm−1 ∞ < 10−5 , where the conditions are
second inequality ensures the stability of the results. When  
[λt ] : pdi,t − pci,t + gbt − gst = Dt − Wt , ∀t ∈ T, (36a)
Algorithm 1 terminates, we only use the resulting integer solu-
tion Z∗ . Then, we solve the MISOCP (25) with x = Z∗ to get 
i∈I

the final continuous solution Y ∗ . [μt ] : adi,t + aci,t = 1, ∀t ∈ T, (36b)
i∈I
 n  n
V. P RICING FOR M ULTIPLE S ERVICES πt : ri,t = rm , ∀n ∈ {d, c}, t ∈ T, (36c)
A. Local Market i∈I

Similar to [20], we employ duality to determine the prices
of the services. By fixing binary variables to x = Z∗ , the
MISOCP (25) becomes a SOCP. Hereafter, the SOCP refers
to the MISOCP (25) with x = Z∗ . Due to the convexity of the
SOCP, we define the service prices as follows.
1) For load shifting, define λt as the multiplier of energy
balance constraint (1). It is the energy price ($/kWh) at
where [ · ] represents the multipliers, namely prices.
B. Market Properties
Here, we theoretically analyze the local market. Given
the prices , both the individual battery problem (34) and
the manager problem (35) are convex programs, for which
Karush–Kuhn–Tucker (KKT) conditions are necessary and

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
ZHONG et al.: CHANCE CONSTRAINED SCHEDULING AND PRICING FOR MULTI-SERVICE BATTERY ENERGY STORAGE 5037

sufficient for optimality. Thus, the conditions (36) and the

KKT conditions of problems (34) and (35) determine the mar-
ket equilibrium. Proposition 1 shows the relation between the
equilibrium and the SOCP.
Proposition 1 (Market Equilibrium): The scheduling result
∗ := {g∗ , p∗ , a∗ , r∗ } (given by the optimal solution to the
SOCP) and the pricing result ∗ := {λ∗ , μ∗ , π ∗ } (given by
the optimal multipliers of the SOCP) constitute a market
equilibrium that satisfies the following properties.
1) Given the prices ∗ , the schedule ∗ provides the Fig. 3. System costs under the proposed PADM-based heuristic (Algorithm 1)
optimal solutions to the individual battery problem (34) and CPLEX (exact algorithm). PFC guarantee: constraint (33) is included in
and the system manager problem (35). Algorithm 1.
2) The schedule ∗ satisfies the market clearing condi-
tions (36).
It is obvious that {∗ , ∗ } corresponds to a point satis-
fying the KKT conditions of the SOCP. One can find that
{∗ , ∗ } also satisfies the KKT conditions of problems (34)
and (35). Thus, ∗ provides the optimal solutions to the two
problems. The detailed proof of Proposition 1 is given in
Appendix B.
In addition, the local market holds two properties: individ-
ual rationality and balanced budget, which are described in
Fig. 4. The committed power rm returned by Algorithm 1 versus the PFC
the following two propositions. The proofs are presented in price Qm . PFC guarantee: constraint (33) is included in Algorithm 1.
Appendix.
Proposition 2 (Individual Rationality): The utility of bat-
tery i, the objective function in (34), is always non-negative prices are set to Qbt = 17.98 e cent/kWh for peak hours t ∈
in the market. {8, . . . , 22} and Qbt = 13.44 e cent/kWh for off-peak hours
Proposition 3 (Balanced Budget): For each service, the t ∈ {1, . . . , 7, 23, 24}. Set Qst = 0.6Qbt . The PFC price is set to
payment made is equal to the payment received in the market Qm = 17.32 e cent/kW that is the average price of frequency
(i.e., no budget deficit or surplus). containment reserve in Dec. 2019 in France [25]. By analyz-
1) For load shifting, the payment made by the load is equal ing the French frequency deviation data in Dec. 2019 [19] and
to the payments received by the grid and batteries: selecting fdb = 10 mHz and fmax = 100 mHz (see Fig. 1),
λ∗t (Dt − Wt ) = Qbt gb∗ we get sampling values Ed = 0.0661 h, Ec = 0.0669 h,
t − Qt gt
s s∗
   σd = 0.0524 h, and σc = 0.0452 h. The parameters of battery
+ λ∗t pd∗
i,t − pi,t , ∀t ∈ T.
c∗
(37) i ∈ {1, 2, 3} are set as follows: Cn2,i = 0.02 e cent/kWh2 ,
i∈I
Cn1,i = 1 e cent/kWh, ηid = ηic = 0.9, Pmax i = 1 MW,
2) For balancing, the payment made by the load is equal {S1B , S2B , S3B } = {4, 6, 8} MWh, Simax = 0.9SiB , Simin = 0.1SiB ,
to the total payment received by the batteries: s1,i = 0.25SiB , and p = s = 0.5. Here, the batteries only
   differ in the capacity parameters, so we can see how capacity
μ∗t = μ∗t ad∗
i,t + ai,t , ∀t ∈ T.
c∗
(38)
impacts the results.
i∈I

3) For PFC, the payment made by the grid is equal to the

B. Computational Results
total payment received by the batteries:
   For the proposed PADM-based heuristic (Algorithm 1), we
Qm rm∗ = πtd∗ ri,t
d∗
+ πtc∗ ri,t
c∗
. (39) 0 = ρ 1 = 1 and set ρ = 10, θ = 0.95.
initialize ρi,t i,t
i∈I t∈T As shown in Fig. 3, the time spent by the proposed heuris-
Equations (37) and (39) imply that the manager objective tic to solve the MISOCP (25) is short. The resulting system
function in (35) is equal to zero at the equilibrium. cost and time spent increase after the PFC guarantee con-
straint (33) is included. The MISOCP (25) can be exactly
solved by CPLEX, but it is greatly time-consuming due to
VI. N UMERICAL R ESULTS
the big problem size. For comparison, we test CPLEX with
A. Parameter Setting time limits and no PFC guarantee. It can be seen in Fig. 3
Set up a system with I = 3 batteries (denoted as B1, B2, and that the proposed solution, [20 s] e130.35, is quite compara-
B3) and a scheduling horizon of T = 24 hours. Load param- ble to the CPLEX solutions, (30 min) e130.78, (1 h) e126.54,
eters (Dt − Wt ) (see Appendix E) are obtained by scaling the (2 h) e125.10, showing that the proposed heuristic can return
hourly consumption and wind generation forecast data from high-quality solutions quickly.
the French grid [19]. Set σD,t = 0.2Dt and σW,t = 0.2Wt [20]. Note that apart from the [30 s] solution, all other solu-
Based on the French electricity tariffs [24], grid energy buy tions in Fig. 3 have zero PFC committed power, namely

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
5038
Fig. 5. Scheduling results. (a) Scheduled discharging power of batteries for load shifting, pd
i,t . (c) Energy traded with the grid, {gt , gt }. (d) Participation factors of batteries for real-time balancing, {ai,t , ai,t }. (e) Discharging power committed by
pc b s
d . (f) Charging power committed by batteries for PFC, rc .
batteries for PFC, ri,t
rm∗ = 0. This means that providing PFC service may not
be a good cost-minimizing choice under our parameter set-
ting. One of the main reasons may be the relatively low
PFC price Qm . To further examine the impact of Qm , we
increase Qm and show the solutions returned by Algorithm 1
in Fig. 4. It is shown that the relation between Qm and
rm is not significant when Qm is relatively low. If Qm is
high enough, the benefit from PFC will attract the system
to increase the committed power. In Fig. 4, the system has
a positive committed power under Qm ≥ 60 e cent/kW
even if the PFC guarantee constraint (33) is not included in
Algorithm 1.
C. Scheduling Results
Fig. 5 shows the scheduling results under the [30 s] solu-
tion in Fig. 3. From Fig. 5(a) and Fig. 5(b), one can see that
B2 and B3 are mostly discharged in peak hours and charged
in off-peak hours. This is an economical decision since the
system can buy less amount of expensive peak-hour energy
from the grid. But, B1 acts a bit differently, discharging in
off-peak hours. This results from the PFC guarantee con-
straint (33) that ensures at least one discharging battery and
one charging battery at each hour. B2 and B3 have higher
capacities than B1, i.e., B2 and B3 can store more energy.
This is the reason why B2 and B3 (rather than B1) are mostly
Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON SMART GRID, VOL. 12, NO. 6, NOVEMBER 2021
i,t . (b) Scheduled charging power of batteries for load shifting,
d c
i,t
charged in off-peak hours. On the operating day, the sched-
uled discharging/charging power will be adjusted to respond
to the realization of uncertainty. Thus, the pattern of par-
ticipation factors in Fig. 5(d) is consistent with the pattern
composed of Fig. 5(a) and Fig. 5(b). For PFC, it can be found
from Fig. 5(e) and Fig. 5(f) that the PFC committed power,
±180.82 kW, is collectively provided by B1, B2, and B3,
satisfying constraints (8).
Fig. 6 shows how the safety parameters {p , s } impact
the scheduling results of B1 under uncertainty realization.
In Fig. 6, the integer solution is fixed and provided by the
[30 s] solution. Each subfigure in Fig. 6 contains 1000 real-
izations of the system uncertainty vector e, in which each
entry is assumed to follow a normal distribution. The power
curves in Fig. 6(a) and Fig. 6(b) are given by pni,t (e) + ri,t
n ,
and the energy curves in Fig. 6(c) and Fig. 6(d) are given
by si,t+1 (e) in (12). It is shown that a smaller safety param-
eter leads to a smaller chance to violate the power/energy
bounds. Also, a smaller safety parameter means a more con-
servative solution, leading to a higher system cost, as shown
in Fig. 7. One can adjust the safety parameters to strike a
balance between the system cost and power/energy bound
violation. In real situations, batteries are not the only type
of resources responding to uncertainty realization. Practical
systems would have other resources such as local generators
ZHONG et al.: CHANCE CONSTRAINED SCHEDULING AND PRICING FOR MULTI-SERVICE BATTERY ENERGY STORAGE 5039

Fig. 6. Actual power and energy state of B1 with safety parameters (a) p = 0.5, (b) p = 0.1, (c) s = 0.5, and (d) s = 0.1. Each gray curve represents
a realization of uncertainty. The power bounds of B1 are given by [0, 1000] kW. The energy bounds of B1 are given by [400, 3600] kWh.

Fig. 7. System cost versus safety parameters (p = s ). PFC guarantee: the
integer solution is fixed and given by the [30 s] solution. No PFC guarantee: Fig. 8. Pricing results. The utility is equal to the incomes minus the cost.
the integer solution is fixed and given by the [20 s] solution.

To further explore the relation between battery utility and

and controllable loads, which can be used to compensate
for the amount of battery power/energy violation on the
operating day.
D. Pricing Results
Fig. 8 shows our pricing results under the [30 s] solution.
It can be seen that each battery has a positive utility, which
validates the property of individual rationality (Proposition 2).
Also, the utilities, incomes, and costs generally go up as the
battery capacity increases. For example, B3 has the highest
capacity, so it can contribute more to service provisioning. As
a result, it is used the most (the highest operating cost) and
earns the most. The pricing results of other solutions in Fig. 3
are similar to that of the [30 s] solution. The major difference
is that the other solutions have zero PFC committed power, so
their PFC incomes are all zero.
capacity, we increase the capacity of B1, S1B , and show the
consequences in Fig. 9, where the energy to power ratio of
B1, S1B /Pmax = 4 h, is kept unchanged. The integer solution
1
is fixed and given by the [30 s] solution. It can be seen that
the utility of B1 grows if B1 capacity increases. The reason
is that B1 can earn more by using more capacity to pro-
vide services. More capacity also leads to less system cost,
as shown in Fig. 9(b). An important observation is that the
rate of system cost reduction decreases as B1 capacity goes
up. This is because we consider quadratic battery usage costs
in (18). This is also reflected in Fig. 9(a), where the rate of
B1 utility growth is slowing down, meaning that the marginal
benefit of capacity is decreasing. Another observation is that
the utility of B3 (8 MWh) is higher than the utility of B1
at 9 MWh in Fig. 9(a). This implies that a higher capacity
does not always lead to a higher utility. But, the capacity
limits how much a battery can be used. In general, under

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
5040 IEEE TRANSACTIONS ON SMART GRID, VOL. 12, NO. 6, NOVEMBER 2021

t < gt . Here, let gt = gt − gt and gt = 0. We

2) gb∗ s∗ s s∗ b∗ b

have Qt gt − Qt gt > Qt (gt − gt ) = −Qst gs

b b∗ s s∗ s b∗ s∗
t .
Both cases imply that {gbt , gt } costs less in trading energy
s

t , gt }. As the MISOCP (25) is a cost-

with the grid than {gb∗ s∗

minimizing problem, a good solution must satisfy gbt gst =0.

n = 0 in (23), then we have pn = zn =
If we fix xi,t i,t i,t
ri,t = yi,t = ani,t = 0. Therefore, any solution meeting
n n

constraints (16) and (23) also meets adi,t aci,t = 0.

A PPENDIX B
P ROOF OF P ROPOSITION 1
Proof: Let Y i denote the decision vector of battery i with
Y i := {pi , ai , ri , wi } where wi := {yi , zi , pri }. Define a convex
function hi (Y i ) such that hi (Y i ) ≤ 0 holds if and only if Y i sat-
isfies the constraints of problem (34). Considering the prices 
are constants, the KKT conditions of battery i’s problem (34)
for all i ∈ I are given by
∂ϕi,t
d
∂hi
− λt + βih = 0, ∀i ∈ I, t ∈ T, (40a)
∂pdi,t ∂pdi,t
∂ϕi,t
c
∂hi
+ λt + βih = 0, ∀i ∈ I, t ∈ T, (40b)
∂pci,t ∂pci,t
∂ϕi,t
n
∂hi
− μt + βih = 0, ∀n ∈ {d, c}, i ∈ I, t ∈ T, (40c)
∂ai,t
n ∂ani,t
Fig. 9. Battery utilities and costs changed with B1 capacity. The capacities ∂ϕi,t
n
∂hi
of B2 and B3 are fixed to 6 and 8 MWh, respectively. The cost of a battery
∂ri,t
n − πtn +βih n = 0, ∀n ∈ {d, c}, i ∈ I, t ∈ T,
∂ri,t
(40d)
reflects how much the battery is used.

∇wi ϕi,t
n
+ βih ∇wi hi = 0, ∀i ∈ I, (40e)
n∈{d,c},
the proposed pricing method, a battery earns more if it is t∈T
 h
used more (see B1); a battery earns less if it is used less βi : hi (Y i ) ≤ 0, ∀i ∈ I, (40f)
(see B2 and B3).
βih ≥ 0, βih hi (Y i ) = 0, ∀i ∈ I. (40g)

VII. C ONCLUSION By the market clearing conditions (36), the objective in (35)
can be written as
In this paper, we proposed day-ahead scheduling and pric-  
ing methods for battery energy storage to provide multiple max Qm rm − rm πtd + πtc
g,rm
services including load shifting, real-time balancing, and PFC. t∈T
    b 
The original stochastic chance-constrained scheduling problem + Qst gst − Qbt gbt + λt gt − gst .
was reformulated to an MISOCP. To obtain suboptimal solu- t∈T t∈T
tions quickly, we developed a PADM-based heuristic to solve
the MISOCP. By fixing the integer solution returned by the Considering the prices  are constants, the KKT conditions
heuristic, we employed the duality of the SOCP to price the of the system manager problem (35) are given by
 
three services in the local market. We also provided theoreti- − Qm + πtd + πtc − βm = 0, (41a)
cal analysis of the market properties. Numerical results show t∈T
that the proposed heuristic is computationally efficient, and
Qbt − λt − βtb = 0, ∀t ∈ T, (41b)
the pricing results can guarantee a positive utility for each bat-
tery, i.e., service incomes are enough to cover battery usage − Qst + λt − βts = 0, ∀t ∈ T, (41c)
costs. [βm ] : rm ≥ 0, (41d)
 b b
β : g ≥ 0, (41e)
 ts  ts
A PPENDIX A βt : gt ≥ 0, (41f)
R EDUNDANCY OF gbt gst = 0 AND adi,t aci,t = 0 βm ≥ 0, βtb ≥ 0, βts ≥ 0, ∀t ∈ T, (41g)
Suppose that there is a solution to the MISOCP (25) with βm rm = 0, βtb gbt = 0, βts gst = 0, ∀t ∈ T. (41h)
t > 0 and gt > 0 for a fixed t. Discuss two cases:
gb∗ s∗

1) gt ≥ gt . In this case, let gb

b∗ s∗
t = gt − gt and gt = 0.
b∗ s∗ s One can find that the KKT conditions (40) and (41) com-
A practical system usually has Qt > Qt > 0, so we get
b s bining with the market clearing conditions (36) are exactly
t − Qt gt > Qt (gt − gt ) = Qt gt .
Qbt gb∗ s s∗ b b∗ s∗ b b the KKT conditions of the SOCP. As {∗ , ∗ } corresponds

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
ZHONG et al.: CHANCE CONSTRAINED SCHEDULING AND PRICING FOR MULTI-SERVICE BATTERY ENERGY STORAGE 5041

to the optimal KKT point of the SOCP, {∗ , ∗ } also A PPENDIX E

satisfies (40), (41), and (36). This completes the proof. D EMAND AND W IND PARAMETERS ( K W H )

A PPENDIX C
P ROOF OF P ROPOSITION 2
Proof: Let u∗i denote the maximum utility of battery i
at the market equilibrium when battery i provides the three
services. Following Proposition 1-1), u∗i can be derived from
the following problem:
    
u∗i = max λ∗t pdi,t − pci,t + μ∗t adi,t + aci,t + πtd∗ ri,t
d
pi ,ai ,ri ,
yi ,zi ,pri t∈T
  n n n n n n
+ πtc∗ ri,t
c
− ϕi,t pi,t , ai,t , ri,t , pi,t ri,t ,
t∈T,
n∈{d,c}
s.t. Constraints in problem (34), (42)

where the prices are fixed to the optimal ∗ . Let u0i denote
the maximum utility of battery i when it does not provide any
service, which means
pni,t = ani,t = ri,t
n
= 0, ∀n ∈ {d, c}, t ∈ T.
n = 0 from (20). We then have u0 = 0
This leads to pni,t ri,t i pp. 2760–2770, Jul. 2020.
according to the utility function in (42). u0i can be also con-
sidered as the optimal objective value of problem (42) with the
additional constraint (43). Thus, we readily have u∗i ≥ u0i = 0,
which means that battery i has a non-negative utility if it
provides the three services. This completes the proof.
This conclusion can be generalized to the situations that
u∗i denotes the maximum utility of battery i when it provides
some of the three services in the market, that is, battery i has
a non-negative utility if it provides at least one of the three
services. For example, if battery i only provides load shifting,
we add the constraint ani,t = ri,t
n = 0 in the MISOCP (25),
the SOCP, and the individual battery problem (34). Then, the
above analysis of problem (42) still holds.
A PPENDIX D
P ROOF OF P ROPOSITION 3
Proof: Load shifting: By Appendix A, the optimal solution
t gt = 0. We discuss three cases:
to the SOCP must have gb∗ s∗
1) gt > 0, gt = 0. In this case, we have βtb∗ = 0 and
b∗ s∗
λ∗t = Qbt according to KKT conditions (41).
∗
t = 0, gt > 0. This case has βt = 0 and λt = Qt .
2) gb∗ s∗ s∗ s microgrids,” IEEE Trans. Smart Grid, vol. 11, no. 4, pp. 3496–3508,
3) gt = gt = 0. This case yields βt , βt > 0 and
b∗ s∗ b∗ s∗
Qst < λ∗t < Qbt .
As per constraint (1), it can be seen that equation (37) holds
in each of the above three cases. This proves Proposition 3-1).
Balancing: The proof of Proposition 3-2) is straightforward.
Equation (38) follows from constraint (6).
PFC: If rm∗ > 0, then we have βm∗ =  0 from KKT con-
ditions (41). This leads to Qm rm∗ = rm∗ t∈T (πtd∗ + πtc∗ ).
Combining with constraints (8), we readily obtain equa-
tion (39). Note that (39) still holds if rm∗ = 0. This completes
the proof of Proposition 3-3).
R EFERENCES
[1] W. Zhong, K. Xie, Y. Liu, C. Yang, and S. Xie, “Multi-resource allo-
cation of shared energy storage: A distributed combinatorial auction
approach,” IEEE Trans. Smart Grid, vol. 11, no. 5, pp. 4105–4115,
(43) Sep. 2020.
[2] N. Y. Soltani and A. Nasiri, “Chance-constrained optimization of energy
storage capacity for microgrids,” IEEE Trans. Smart Grid, vol. 11, no. 4,
[3] W. Zhong, K. Xie, Y. Liu, C. Yang, S. Xie, and Y. Zhang, “Online
control and near-optimal algorithm for distributed energy storage sharing
in smart grid,” IEEE Trans. Smart Grid, vol. 11, no. 3, pp. 2552–2562,
May 2020.
[4] R. Hollinger, A. M. Cortes, and T. Erge, “Fast frequency response with
BESS: A comparative analysis of Germany, Great Britain and Sweden,”
in Proc. EEM, Lodz, Poland, 2018, pp. 1–6.
[5] D. Wu, C. Jin, P. Balducci, and M. Kintner-Meyer, “An energy stor-
age assessment: Using optimal control strategies to capture multiple
services,” in Proc. IEEE Power Energy Soc. Gener. Meeting, Denver,
CO, USA, 2015, pp. 1–5.
[6] O. Mégel, J. L. Mathieu, and G. Andersson, “Scheduling distributed
energy storage units to provide multiple services under forecast error,”
Elect. Power Energy Syst., vol. 72, pp. 48–57, Nov. 2015.
[7] R. Moreno, R. Moreira, and G. Strbac, “A MILP model for optimising
multi-service portfolios of distributed energy storage,” Appl. Energy,
vol. 137, pp. 554–566, Jan. 2015.
[8] E. Namor, F. Sossan, R. Cherkaoui, and M. Paolone, “Control of battery
storage systems for the simultaneous provision of multiple services,”
IEEE Trans. Smart Grid, vol. 10, no. 3, pp. 2799–2808, May 2019.
[9] S. Gupta, V. Kekatos, and W. Saad, “Optimal real-time coordination of
energy storage units as a voltage-constrained game,” IEEE Trans. Smart
Grid, vol. 10, no. 4, pp. 3883–3894, Jul. 2019.
[10] R. A. Jabr, S. Karaki, and J. A. Korbane, “Robust multi-period OPF
with storage and renewables,” IEEE Trans. Power Syst., vol. 30, no. 5,
pp. 2790–2799, Sep. 2015.
[11] Q. Xu, T. Zhao, Y. Xu, Z. Xu, P. Wang, and F. Blaabjerg, “A distributed
and robust energy management system for networked hybrid AC/DC
Jul. 2020.
[12] A. Kargarian, G. Hug, and J. Mohammadi, “A multi-time scale co-
optimization method for sizing of energy storage and fast-ramping
generation,” IEEE Trans. Sustain. Energy, vol. 7, no. 4, pp. 1351–1361,
Oct. 2016.
[13] Y. Li, Z. Yang, G. Li, D. Zhao, and W. Tian, “Optimal scheduling of an
isolated microgrid with battery storage considering load and renewable
generation uncertainties,” IEEE Trans. Ind. Electron., vol. 66, no. 2,
pp. 1565–1575, Feb. 2019.
[14] Z. Shi, H. Liang, S. Huang, and V. Dinavahi, “Distributionally robust
chance-constrained energy management for islanded microgrids,” IEEE
Trans. Smart Grid, vol. 10, no. 2, pp. 2234–2244, Mar. 2019.
[15] W. Xie and S. Ahmed, “Distributionally robust chance constrained
optimal power flow with renewables: A conic reformulation,” IEEE
Trans. Power Syst., vol. 33, no. 2, pp. 1860–1867, Mar. 2018.

Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
5042
[16] B. Geißler, A. Morsi, L. Schewe, and M. Schmidt, “Penalty alternat-
ing direction methods for mixed-integer optimization: A new view on
feasibility pumps,” SIAM J. Optim., vol. 27, no. 3, pp. 1611–1636, 2017.
[17] J. Kang, Z. Xiong, D. Niyato, S. Xie, and D. I. Kim, “Securing data
sharing from the sky: Integrating blockchains into drones in 5G and
beyond,” IEEE Netw., vol. 35, no. 1, pp. 78–85, Jan./Feb. 2021.
[18] X. Huang, D. Ye, R. Yu, and L. Shu, “Securing parked vehicle assisted
fog computing with blockchain and optimal smart contract design,”
IEEE/CAA J. Autom. Sinica, vol. 7, no. 2, pp. 426–441, Mar. 2020.
[19] (Feb. 2021). Réseau de Transport d’Électricité (RTE). [Online].
Available: https://www.services-rte.com/en/home.html
[20] Y. Dvorkin, “A chance-constrained stochastic electricity market,” IEEE
Trans. Power Syst., vol. 35, no. 4, pp. 2993–3003, Jul. 2020.
[21] W. Zhong, S. Xie, K. Xie, Q. Yang, and L. Xie, “Cooperative P2P energy
trading in active distribution networks: An MILP-based Nash bargain-
ing solution,” IEEE Trans. Smart Grid, vol. 12, no. 2, pp. 1264–1276,
Mar. 2021.
[22] G. P. McCormick, “Computability of global solutions to factorable non-
convex programs: Part I—Convex underestimating problems,” Math.
Prog., vol. 10, no. 1, pp. 147–175, 1976.
[23] J. Kronqvist, D. E. Bernal, A. Lundell, and I. E. Grossmann, “A review
and comparison of solvers for convex MINLP,” Optim. Eng., vol. 20,
no. 2, pp. 397–455, 2019.
[24] (Feb. 2021). Tarif Bleu: Regulated Sale Tariff for Electricity.
[Online]. Available: https://particulier.edf.fr/en/home/energy-and-
services/electricity/tarif-bleu.html
[25] (Feb. 2021). Regelleistung. [Online]. Available: https://www.
regelleistung.net
Weifeng Zhong received the Ph.D. degree in control
science and engineering from Guangdong University
of Technology, Guangzhou, China, in 2019, where
he is currently an Associate Professor. He was
a Visiting Scholar with Nanyang Technological
University, Singapore, in 2021, and a visiting stu-
dent with the Hong Kong University of Science
and Technology, Hong Kong, in 2016. His research
interest lies in optimization, resource allocation, and
economics in smart grid.
Kan Xie received the Ph.D. degree in control science
and engineering from Guangdong University of
Technology, Guangzhou, China, in 2017. He joined
the Institute of Intelligent Information Processing,
Guangdong University of Technology, where he
is currently an Associate Professor. His research
interests include machine learning, nonnegative sig-
nal processing, blind signal processing, smart grid,
and Internet of Things.
Authorized licensed use limited to: University Town Library of Shenzhen. Downloaded on August 15,2022 at 09:53:13 UTC from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON SMART GRID, VOL. 12, NO. 6, NOVEMBER 2021
Yi Liu received the Ph.D. degree in signal and
information processing from South China University
of Technology, Guangzhou, China, in 2011. He
joined the Singapore University of Technology and
Design as a Postdoctoral Researcher. In 2014, he
worked with the Institute of Intelligent Information
Processing, Guangdong University of Technology,
where he is currently a Full Professor. His
research interests include wireless communication
networks, cooperative communications, smart grid,
and intelligent edge computing.
Shengli Xie (Fellow, IEEE) received the B.S. degree
in mathematics from Jilin University, Changchun,
China, in 1983, the M.S. degree in mathematics from
Central China Normal University, Wuhan, China,
in 1995, and the Ph.D. degree in control theory
and applications from South China University of
Technology, Guangzhou, China, in 1997. He is cur-
rently a Full Professor and the Head of the Institute
of Intelligent Information Processing, Guangdong
University of Technology, Guangzhou. He has coau-
thored two books and more than 150 research papers
in refereed journals and conference proceedings and was awarded Highly
Cited Researcher in 2020. His research interests include blind signal process-
ing, machine learning, and Internet of Things. He was awarded the Second
Prize of National Natural Science Award of China in 2009. He is an Associate
Editor for IEEE T RANSACTIONS ON S YSTEMS , M AN , AND C YBERNETICS :
S YSTEMS.
Lihua Xie (Fellow, IEEE) received the Ph.D. degree
in electrical engineering from the University of
Newcastle, Australia, in 1992.
Since 1992, he has been with the School of
Electrical and Electronic Engineering, Nanyang
Technological University, Singapore, where he is
currently a Professor and the Director of Delta-NTU
Corporate Laboratory for Cyber-Physical Systems
and the Center for Advanced Robotics Technology
Innovation. He served as the Head of Division of
Control and Instrumentation from July 2011 to June
2014. He held teaching appointments with the Department of Automatic
Control, Nanjing University of Science and Technology from 1986 to
1989. His research interests include robust control and estimation, net-
worked control systems, multiagent networks, localization, and unmanned
systems. He is an Editor-in-Chief for Unmanned Systems and has served
as an Editor of IET Book Series in Control and an Associate Editor of
a number of journals, including IEEE T RANSACTIONS ON AUTOMATIC
C ONTROL, Automatica, IEEE T RANSACTIONS ON C ONTROL S YSTEMS
T ECHNOLOGY, IEEE T RANSACTIONS ON N ETWORK C ONTROL S YSTEMS,
and IEEE T RANSACTIONS ON C IRCUITS AND S YSTEMS -II. He was an IEEE
Distinguished Lecturer from January 2012 to December 2014. He is a Fellow
of Academy of Engineering Singapore, IFAC, and CAA.